Two Nature papers describe multi-agent AI 'co-scientists' from Google DeepMind and FutureHouse
Co-Scientist and Robin, published in Nature on 19 May, use teams of AI agents to propose and rank hypotheses. Both report early cell-based lab results and both say the work is preliminary.

On 19 May 2026, Nature published two papers describing multi-agent AI systems built to help scientists generate, rank and refine research hypotheses. Co-Scientist comes from Google DeepMind, Google Research and Google Cloud, working with academic collaborators. Robin comes from FutureHouse. Both papers report laboratory follow-up of ideas the systems proposed, and both describe that evidence as early.
The two systems were first described in arXiv preprints in February and May 2025 respectively. Peer-reviewed publication with wet-lab results is the news here, and a useful snapshot of where AI hypothesis generation stands.
Two systems, one shape
The papers were developed independently, but the designs have a lot in common. Each splits the work across specialised agents, grounds ideas in the published literature, and ranks competing hypotheses with a tournament judged by a language model.
| Co-Scientist | Robin | |
|---|---|---|
| Developer | Google DeepMind, Google Research, Google Cloud | FutureHouse |
| Base models (as reported) | Gemini 2.0 for all agents | OpenAI o4-mini for literature work and hypotheses; Claude 3.7 Sonnet as judge |
| Ranking method | Elo tournament of pairwise “debates” | LLM-judged pairwise tournament |
| Scope | General purpose; validated in biomedicine | Experimental biology; applied to one disease |
Co-Scientist uses a supervisor agent that turns a research goal into tasks and allocates them to agents for generation, reflection (acting as a peer reviewer), ranking, proximity (clustering similar ideas), evolution (refining the best ones) and meta-review. It improves through feedback written back into prompts rather than through retraining. The authors write that automated evaluations “show continued benefits of test-time compute scaling”, meaning hypotheses rated higher the longer the system ran. They also note that this Elo metric “is auto-evaluated and not based on independent ground truth”.
Robin pairs two literature-search agents, Crow and Falcon, with a data-analysis agent called Finch that writes and runs code in Jupyter notebooks. That lets it propose an experiment, read the resulting data and suggest the next step.
What the labs found
Co-Scientist was tested on three biomedical problems:
- Drug repurposing for acute myeloid leukaemia (AML). Of five candidate drugs tested in cell lines, the paper reports that binimetinib, pacritinib and cerivastatin inhibited cell viability. A further candidate, the IRE1α inhibitor KIRA6, also inhibited viability.
- Liver fibrosis. The system proposed three epigenetic targets and drugs acting on them. In human hepatic organoids, two “exhibited significant anti-fibrotic activity” without causing cell toxicity. One of them, vorinostat, is already approved by the US Food and Drug Administration for a cancer indication.
- Antimicrobial resistance. On a question about how a family of mobile genetic elements spreads between bacteria, the system’s top-ranked hypothesis (that they interact with diverse phage tails to widen the range of hosts they can reach) matched the main finding of an independent study that had not yet completed peer review. The authors say the idea was “generated by the AI in just 2 days”.
Robin was applied to dry age-related macular degeneration, which the Nature press release describes as a major cause of blindness in developed countries. It proposed boosting phagocytosis, the process by which cells in the retinal pigment epithelium engulf and clear material, picked an assay, and ranked candidate drugs. In cell experiments, ripasudil, a clinically used Rho kinase inhibitor, increased phagocytosis. The authors say that to their knowledge it “has never previously been proposed” for this disease. Robin then proposed an RNA sequencing follow-up experiment; according to the press release, its suggested follow-up studies identified further potential drug targets. As a comparison, the team gave the same starting prompt to OpenAI’s Deep Research; the paper reports that none of the drugs it suggested were hits in the assay.
In Robin’s case, humans chose which ranked candidates to test, wrote the lab protocols and ran the experiments; the paper says all hypotheses, experimental directions, analyses and main figures were produced by the system.
Stated limitations
Both teams are explicit about the gaps.
- Co-Scientist’s knowledge “is constrained by its reliance on open-access scientific literature”, so it can miss paywalled work and negative results. The authors warn of “a risk of propagating erroneous or irreproducible findings” and of “the potential for hallucinations”. They describe the validation as “preliminary” and say the cell experiments are “not a replacement for the rigorous preclinical and clinical assessment” needed for any therapy.
- Robin “does not yet produce precise, executable protocols”, and its analysis agent is “reliant on prompt engineering by domain experts”. The authors say in vivo validation of the drugs would be needed.
Neither paper claims a new treatment. All of the reported lab tests were in cell lines or organoids.
Access
Google DeepMind said that an experimental tool for individual researchers, called Hypothesis Generation, would roll out “in the coming weeks” via a sign-up at labs.google/science, with an enterprise preview available through Google Cloud. The company describes Co-Scientist as “a partner in research, not a replacement for scientific or clinical expertise”. FutureHouse has published the code for Robin and its Finch analysis agent on GitHub.
What it means
For people building agent systems, the convergence is the interesting part. Two independent teams arrived at a similar recipe: separate generation from critique, ground claims in retrieved literature, and use pairwise tournaments with a model as judge to sort a large pool of ideas. Co-Scientist’s results also suggest that spending more compute at inference time keeps improving the system’s own quality scores, although those scores are self-assessed.
The scientific evidence is narrower than the architecture story. Each system was validated on a handful of problems, with lab work limited to cell lines and organoids, and with humans choosing what to test and running the experiments. That is a reasonable first bar for a hypothesis generator, but medicine’s bar is whether ideas survive animal studies and clinical trials. The next useful signal will be independent groups using these tools on their own problems and reporting the misses as well as the hits.
Sources
- Nature: Accelerating scientific discovery with Co-Scientist (Gottweis et al., 2026)
- Nature: A multi-agent system for automating scientific discovery (Ghareeb et al., 2026)
- Nature press release via Nature Asia: AI research assistants that may accelerate scientific discovery
- Google DeepMind: Co-Scientist, a multi-agent AI partner to accelerate research
- FutureHouse Robin on GitHub



